As DARROW enters its final year, we are not only highlighting the project’s technical results, but also the people behind them. Throughout this interview series, DARROW partners reflect on the project’s outcomes, the challenges they faced, and the lessons learned along the way.
In this interview, Omid Sobhani shares his perspective as a PhD researcher at imec, working at the intersection of artificial intelligence and industrial process systems. He reflects on the development of reinforcement learning for wastewater treatment control and what it takes to bring AI into real industrial practice.
Tell us a bit about yourself and your organization. What role have you played in DARROW?
My name is Omid Sobhani. I am a PhD researcher at imec, working at the intersection of artificial intelligence and industrial process systems. My research focuses strongly on reinforcement learning for real-world, safety-critical applications. I am motivated by closing the gap between advanced AI methods and what is feasible and reliable in industrial environments.
Within the DARROW project, I contributed as a research partner specialising in AI. I was involved in developing and evaluating reinforcement-learning-based approaches for control and optimisation under realistic operational constraints. This included working with data-driven models, and designing objective and constraint formulations that reflect real industrial priorities rather than idealised benchmarks.
Importantly, this work was carried out in close collaboration with Ghent University. Together, we aligned AI methodology with process-level insights, validated modelling and control assumptions, and ensured that the learning setup remained physically meaningful and practically relevant. This collaborative approach was essential for grounding the AI development in real system behaviour and industrial use cases.
In your view, what is the most valuable innovation or tool that DARROW has developed?
In my view, the most valuable innovation developed in DARROW is the autonomous reinforcement learning controller. Unlike traditional control methods, this agent learns by interacting with a high-fidelity simulation environment that is grounded in historical data and shaped by deep system knowledge. This setup allows the controller to dynamically adapt its strategy to changing process conditions while strictly adhering to real-world industrial constraints, such as safety limits and efficiency targets. By combining data-driven learning with physical process insights, we have created a robust system capable of handling routine decision-making, allowing human operators to focus on high-level oversight and complex problem-solving.
If you had to describe DARROW in one sentence, what would it be?
DARROW leverages AI to make wastewater treatment processes smarter and more adaptive, thereby improving operational efficiency, safety, and reliability.
What is something you have contributed to DARROW that you are especially proud of?
I am particularly proud of my contribution to the development and implementation of the adaptive AI controller for wastewater treatment. I led the work on designing how the system can respond dynamically to different operating conditions, balancing multiple objectives like treatment efficiency, safety, and operational stability.
This was challenging because the process is highly non-linear and subject to unexpected changes. However, by combining process knowledge with historical data, we created a controller that can make intelligent, context-aware decisions. It was very rewarding to see it perform reliably in simulation and to know that it has the potential to guide or even handle real operations.
What was one of the biggest challenges you or your team faced, and how did you overcome it?
One of the biggest challenges we faced in DARROW was bridging the gap between AI research and real industrial processes. Industrial operations come with complex constraints, safety requirements, and practical considerations that are not always easily translated into information that an AI algorithm can learn from.
Our approach was twofold. First, we translated industrial objectives and constraints into a format that the AI controller could understand and optimise. Second, we designed the AI output to be interpretable and actionable for human operators. This iterative process of translation, combined with the close collaboration with Ghent University and process engineers, allowed us to develop a system that is both technically sound and practically relevant.
What advice would you give to a future project team taking on something as ambitious as DARROW?
My advice would be to think big, but start small. DARROW tackles complex, unpredictable systems, so it is easy to get lost in theory. Instead, build something that works in the messy, real world first, even if it is imperfect, and allow it to evolve.
Also, embrace the unexpected. The system will behave differently to how your simulations predict. It is from these differences that you will learn the most valuable lessons. Treat every surprise as an opportunity to learn, refine, and improve.
Finally, collaborate relentlessly. The magic of DARROW is not just the AI alone; it is in the conversation between engineers, operators, and researchers. The team that listens, adapts, and iterates together will turn ambitious ideas into solutions that actually work.
What kind of change do you think DARROW can bring to how we manage water and resources in the future?
I believe that DARROW has the power to change the way we think about water and resource management by shifting it from reactive to anticipatory and adaptive. Instead of operators responding to problems after they occur, AI-driven systems can continuously monitor, predict, and adjust operations in real time, thereby optimising efficiency, safety, and sustainability.
This could transform treatment plants into intelligent ecosystems, where water, energy, and chemicals are dynamically managed to minimise waste and environmental impact. Ultimately, projects like DARROW could make water management more resilient and intelligent, enabling us to handle unexpected challenges more effectively while keeping systems safe and reliable.
Omid highlights a central ambition of DARROW: not just to experiment with AI, but to embed it meaningfully within operational reality. By translating industrial constraints into learnable structures and designing systems that remain interpretable and controllable, the project demonstrates how advanced algorithms can move from theory to trusted tools.